When someone asks for an “AI summary” of a dashboard, they usually want a diagnosis and a prescription. They do not want a prettier restatement of what already moved week-on-week or month-on-month.
At WPP Media I sat in analytics and MarTech. Marketers already had the table. What they asked for was what happened, and what to do next. Pasting that table into a model does not change the ask. It only changes the wrapper.
A week-on-week or month-on-month move is a useful signal. It is not decision-making. “ROAS is down” is a starting line. If the summary stops there, you have a caption for a chart.
What useful actually looks like
A useful summary leads with the recommendation. Then it lists the reasons: the potential drivers that put that action on the table.
First, the prescription. A next step a person can run. Not “watch ROAS.”
Second, the reasons. Not one confident story. A short list of candidate drivers that could explain the move, so the person can see why that recommendation is the one being made.
That is diagnosis plus prescription, in that working order. A descriptive rewrite of the table is not that work.
The dashboard example
Take the line everyone already has: ROAS is down.
A descriptive summary restates the percentage and maybe names the worst campaign.
A useful one starts with the action. If search ads had better ROAS than social, the recommendation is a budget shuffle: move daily budget from Meta Ads to Google Ads / SEM. That transfer is the point of the summary.
Then the reasons sit under that recommendation as a diagnostic list. Auction pressure may have thinned delivery. A creative set may have aged out. Targeting may have narrowed. Daily budget may already have moved from campaign A to campaign B, or from Meta Ads to Google Ads, when SEM started to look cheaper on the day. Those are potential drivers. They explain why the shuffle is in play.
The first brief is usually too thin
The first brief is often a performance export. Media and performance numbers are necessary. They are rarely enough.
You also need targeting and creative. UTM parameters can encode which creative is used, so that trail is often already in the tracking if you pass it in. And you especially need a view of daily budget allocation — the shuffles at campaign or ad-set level. Those allocation decisions are often not recorded as a structured “reason” field in a performance-only file. The model sees that spend moved from A to B, or from Meta Ads to Google Ads when search outperformed social. It does not see why the money moved.
That is budget shuffle, not a promo discount. If you only give the model the result table, it will narrate the result. It cannot reconstruct a decision log you never exported.
A chart is not a business
Even that stack is still not enough.
A summary that does not know the business model — how the business earns money — will advise the chart. It will tell you which bar moved. It will not tell you whether that move matters for a retailer that makes money on repeat purchase, a bank that makes money on approved applications, or a brand that is buying reach on purpose.
The operating model is a different thing. It is how the organisation runs the work: who can move budget, how often allocation is reviewed, what the team is allowed to change this week. Mixing those two up is how a summary starts talking about process when the ask was commercial.
I wrote earlier that AI adoption is still a translation problem — people need to know what an answer is for. I also showed that TabFM could forecast the KPI and still could not answer the planner’s budget question. The same gap shows up here. A forecast of a chart is still a chart answer. The planner’s question is commercial: given how this business earns money, where should spend go? A fluent recap of a dashboard that does not know the business model is still a recap.
What I would ask for
If you want the summary to be useful, brief the model the way you would brief an analyst you trust.
Give it the performance. Give it targeting and creative, including the UTM trail that can encode which creative ran. Give it the daily allocation view, including the shuffles. Then give it the business model — how money is made. Give it the operating model only when the team needs to know what it can actually change this week.
Ask for the recommendation first, then the potential drivers that support it. Keep a person on the recommendation. The model can draft the pack. The decision is still yours.
When you ask for an AI summary of a dashboard, which of those pieces is usually missing from the brief? One way to address that is RAG: retrieve the missing context into the summary brief, instead of hoping the performance export already contains it.
